Git hooks turned into a continuous AI review pipeline

What it does

roborev is a Go CLI that hijacks your git post-commit hook to run a local, continuous review queue. After every commit, it ships the diff to an AI coding agent you already have installed—Claude Code, Codex, Gemini, Copilot, Cursor, among others—and surfaces findings in a terminal UI or on the command line. It also runs targeted analyses for duplication, complexity, dead code, security, and more, optionally handing results back to the agent for repair.

The interesting bit

Instead of waiting for a pull request to batch feedback, roborev keeps the review loop hot while context is fresh. It can park an agent in an isolated worktree and iterate through fix and re-review cycles until the code passes or a ceiling is hit—an attempt to stop LLM-generated technical debt from compounding in real time.

Key highlights

• Runs entirely locally with no hosted service or additional infrastructure.
• Supports a broad roster of agents (Codex, Claude Code, Gemini, Copilot, Cursor, Kiro, Kilo, Droid, Pi, OpenCode) and auto-detects installed ones.
• Built-in analysis modes for duplication, complexity, refactoring, test fixtures, dead code, API design, architecture, and security.
• compact verifies open findings against current code to filter false positives and consolidate related issues.
• Can route Claude Code through local proxies like Ollama or LiteLLM with per-task model specs and explicit token handling.

Caveats

• The README explicitly warns that roborev is built for trusted codebases: review and fix agents receive shell access, and untrusted diffs should be sandboxed to limit prompt-injection blast radius.
• The fully automated refine loop is labeled an advanced feature; it repeatedly spins up worktrees and re-reviews, which implies meaningful local compute and time costs.

Verdict

A solid fit if you already live inside an AI agent workflow and want a local, commit-level safety net. Less appealing if you prefer centralized CI gates, or if the idea of background LLM processes with shell access in your repo makes you nervous.